Bitcoin is not the New Gold – A comparison of volatility, correlation,and portfolio performance

Klein, T., Thu, H. P., & Walther, T. (2018). Bitcoin is not the New Gold – A comparison of volatility, correlation,and portfolio performance. International Review of Financial Analysis, 59, 105-116.https://doi.org/10.1016/j.irfa.2018.07.010

Published in:International Review of Financial Analysis

Document Version:Peer reviewed version

Queen's University Belfast - Research Portal:Link to publication record in Queen's University Belfast Research Portal

Publisher rightsCopyright 2018 Elsevier.This manuscript is distributed under a Creative Commons Attribution-NonCommercial-NoDerivs License(https://creativecommons.org/licenses/by-nc-nd/4.0/), which permits distribution and reproduction for non-commercial purposes, provided theauthor and source are cited

General rightsCopyright for the publications made accessible via the Queen's University Belfast Research Portal is retained by the author(s) and / or othercopyright owners and it is a condition of accessing these publications that users recognise and abide by the legal requirements associatedwith these rights.

Take down policyThe Research Portal is Queen's institutional repository that provides access to Queen's research output. Every effort has been made toensure that content in the Research Portal does not infringe any person's rights, or applicable UK laws. If you discover content in theResearch Portal that you believe breaches copyright or violates any law, please contact openaccess@qub.ac.uk.

Download date:10. Dec. 2021

Bitcoin is not the New Gold – A Comparison of Volatility,

Correlation, and Portfolio PerformanceI

Tony Kleina,b,∗, Hien Pham Thuc, Thomas Walthera,d

aFaculty of Business and Economics, Technische Universitat Dresden, GermanybQueen’s Management School, Queen’s University Belfast, UK

cSchool of Business and Economics, Humboldt-Universitat zu Berlin, GermanydInstitute for Operations Research and Computational Finance, University of St. Gallen, Switzerland

Abstract

Cryptocurrencies such as Bitcoin are establishing themselves as an investment asset and

are often named the New Gold. This study, however, shows that the two assets could

barely be more different. Firstly, we analyze and compare conditional variance proper-

ties of Bitcoin and Gold as well as other assets and find differences in their structure.

Secondly, we implement a BEKK-GARCH model to estimate time-varying conditional

correlations. Gold plays an important role in financial markets with flight-to-quality in

times of market distress. Our results show that Bitcoin behaves as the exact opposite

and it positively correlates with downward markets. Lastly, we analyze the properties of

Bitcoin as portfolio component and find no evidence for stable hedging capabilities. We

conclude that Bitcoin and Gold feature fundamentally different properties as assets and

linkages to equity markets. Our results hold for the broad cryptocurrency index CRIX. As

of now, Bitcoin does not reflect any distinctive properties of Gold other than asymmetric

response in variance.

Keywords: Bitcoin, Correlation, Cryptocurrencies, Gold, Asymmetry

JEL classification: C10; C58; G11

1. Introduction

The popularity of cryptocurrencies has risen significantly since Nakamoto (2008) in-

troduced the concept of Bitcoin. Cryptocurrencies embody innovative technology, high-

security architecture, prosperity in functionalities, and investment opportunity as an as-

set whats makes them attractive for computer scientists, venture capitalists as well as

investors. However, the decentralization and unregulated markets add an additional layer

of uncertainty to its pricing and projection of application. Examples are the closures

IWe thank Wolfgang Hardle and Hermann Locarek-Junge and two anonymous reviewers for theirhelpful remarks. We are also thankful to the participants of the “Cryptocurrencies in a Digital Economy”workshop (HU Berlin), in particular Simon Trimborn, for their valuable suggestions and comments.

∗Corresponding Author, Mail: t.klein@qub.ac.uk.

Working Paper Version 2 - May 21, 2018 - Page 1

of exchanges in China based on changing legal situation which causes worldwide price

reactions of large magnitude. Especially in the last years, large shocks and a bubble-like

price movement are observable. The market capitalization of cryptocurrencies reached its

peak at USD 831bn in January 7, 2018 which is comparable to the market valuation of

the biggest companies in the world, like Apple or Alphabet. The market with currently

1 622 cryptocurrencies (as of May 2018) is dominated by Bitcoin, Ethereum, and Ripple

which capture around 63 percent of the total market capitalization.1 What started as an

experiment by decentralized governance enthusiasts is becoming an investment vehicle for

investors and a core business model for companies as a new and innovative way of pay-

ment. The general idea of this distributed ledger technology is transferred to many other

fields in finance. For a detailed introduction into the key concepts of cryptocurrencies

and Bitcoin, we refer to Bariviera et al. (2017), Chuen (2015), Dwyer (2015), Elendner

et al. (2018), and Hardle et al. (2018).

Cryptocurrencies, in particular Bitcoin, have been labeled the New Gold by some me-

dia, banks, and also data providers throughout the last years.2 While this view might

be motivated by fast and high returns in a gold rush like environment, we compare Gold

and Bitcoin from an econometric perspective and focus on the economic aspects of cryp-

tocurrencies as an investment asset. We address the question how cryptocurrencies can

be classified based on volatility behavior and how they are correlated with already es-

tablished asset classes. Thus, we do not explore the cryptocurrency market itself as in

Brandvold et al. (2015) or Ciaian et al. (2017), but the relationship with other asset

classes. Cryptocurrencies are not directly linked to any monetary policy instruments or

fundamentals. Therefore, analysis of common factors between these virtual currencies

and other financial asset classes is challenging. On the regulatory side, the Commodity

Futures Trading Commission (CFTC) has officially declared virtual money a commodity,

just like crude oil or Gold. The Commission states that Bitcoin as a virtual currency is a

digital representation of value that functions as a medium of exchange, a unit of account,

and/or a store of value, but does not have legal tender status in any jurisdiction. Bitcoin

and other cryptocurrencies are distinctly different from fiat currencies, which are the coin

and paper money that are designated as legal tender, circulated, and are customarily used

and accepted as a medium of exchange.

The analysis in this paper is divided into three parts. Firstly, we investigate the

volatility behavior of cryptocurrencies in comparison to stock indices and commodities.

According to Blau (2017), the high volatility of Bitcoin until 2014 is not related to specu-

lative trading. This is in contradiction to Cheah & Fry (2015) and Cheung et al. (2015),

who find Bitcoin to build speculative bubbles in the same time frame. In the present study,

1The calculations are based on data from www.coinmarketcap.com retrieved May 16, 2018.2See for example articles on Bloomberg, Forbes, and CNN.

2

we aim to classify cryptocurrencies within the conventional asset classes by analyzing the

volatility. Conrad et al. (2018), Catania & Grassi (2017), Dyhrberg (2016), Katsiampa

(2017), and Chu et al. (2017) use univariate volatility models to analyze the volatility

structure of Bitcoin and other cryptocurrencies. Conrad et al. (2018) examine the drivers

of long-term volatility of Bitcoin and compare them to other assets class, e.g. Gold. The

authors find that Bitcoin volatility is distinct compared to other asset classes. Here, we

focus especially on the stylized facts long memory and asymmetry. The property of long

memory of a financial time series is also referred to as persistence and describes long last-

ing, i.e. slowly decaying autocorrelation effects in conditional returns or volatility (Baillie,

1996). Many types of financial time series are reported to attribute long memory in their

variance; e.g. individual stocks, stock indices, commodities, and foreign exchange rates

(i.a. Baillie, 1996, Bollerslev & Mikkelsen, 1996, Chkili et al., 2014, Walther et al., 2017).

The property of asymmetric volatility is explained by the empirical phenomenon when

negative (positive) returns are associated with upward (downward) revision of the condi-

tional volatility (Engle & Ng, 1993, Zakoian, 1994). In the stock market, this asymmetry

effect results from the leverage effect (Black, 1976, Christie, 1982) and the volatility feed-

back effect (Campbell & Hentschel, 1992). Regarding cryptocurrencies, Dyhrberg (2016)

reports an insignificant leverage effect of Bitcoin. Catania & Grassi (2017), however,

find an inverse leverage effect as well as significant long memory in the most prominent

cryptocurrencies.

Secondly, this research explores the hedge and safe haven capabilities of cryptocur-

rencies in comparison to Gold by means of a dynamic correlation analysis. We apply the

definition of hedge, diversifier, and safe haven given in Baur & Lucey (2010). An asset

which is uncorrelated or negatively correlated with another asset is defined as a hedge

whereas a safe haven asset is uncorrelated or negatively correlated with other assets in

distressed markets only. Assets which are a diversifier are positively (on average), but

not perfectly correlated to other assets. As diversification opportunities are diminishing

in market turmoil, investors seek to find assets which are uncorrelated or negatively cor-

related with their portfolio’s assets. There is a wide range of literature investigating Gold

as hedge and safe haven against assets such as stocks, bonds, and US Dollar (i.a. Baur

& Lucey, 2010, Baur & McDermott, 2010, 2016, Capie et al., 2005). However, on more

recent data, the safe haven property seems to be dissipating (Klein, 2017). Meanwhile,

research on cryptocurrencies’ purpose for investment is growing. For example, Dyhrberg

(2016) compares the hedging capabilities of Bitcoin and Gold against stocks and US Dol-

lar. Bouri et al. (2017b) conclude that Bitcoin may only be used as a diversifier, but

not as a hedge. However, Bouri et al. (2017a) and Demir et al. (2018) note that Bitcoin

can be used as short-term hedge in extreme market situations. The results of Corbet

et al. (2018) suggest that cryptocurrencies are not connected to conventional markets and

may offer short-term diversification benefits. Guesmi et al. (2018) finds short positions

3

in Bitcoin to have hedging potential. We discuss short positions at a later point. In our

analysis, we assess the potential of cryptocurrencies as hedge and safe haven. We examine

the correlation of Bitcoin with Gold, Silver, the oil price WTI, and the three equity in-

dices S&P 500, MSCI World, and MSCI Emerging Markets 50. The diversification effects

provided by cryptocurrencies are investigated in the context of time-varying correlation.

Lastly, we apply a portfolio analysis which emphasizes the behavior of Gold and Bitcoin

in times of distress.

Our findings indicate that Bitcoin is not the new Gold. Its volatility dynamics share

some aspects with Gold and Silver, however, from a portfolio perspective, Bitcoin does

not serve as a safe-haven which is a prominent feature of Gold.

The structure of this paper is organized as follows. Section 2 presents the methodology

of the study in which we outline the comparison based on univariate volatility modeling,

multivariate variance-covariance modeling, and from a portfolio perspective. The data

sets and preliminary tests are introduced in Section 3. Section 4 presents the findings of

the comparison and discusses them. The paper concludes in Section 5.

2. Methodology

2.1. Properties of Conditional Variance

In order to characterize the volatility structure of assets, we employ (Generalized)

Autoregressive Conditional Heteroskedasticity ((G)ARCH, Engle, 1982, Bollerslev, 1986)

models. Here, we focus especially on two stylized facts: the leverage effect and long mem-

ory. Thus, for our volatility regression we use two models which are able to depict these

two properties. Namely, we use the Asymmetric Power ARCH (APARCH, Ding et al.,

1993) and the Fractionally Integrated APARCH (FIAPARCH, Tse, 1998) model. For-

mally, we run a first order autoregressive model on the asset’s returns, rt, with Student’s

t distributed errors of zero mean with conditional variance ht which then reads

rt = θ0 + θ1rt−1 + εt,

εt =√htηt,

with ηt ∼ St-tν(0, 1) i.i.d. for all t = 1, . . . , n. Table 1 summarizes the definitions of ht for

both models. The non-negative constraints ω, β ≥ 0 have to hold. Further, FIAPARCH

requires 0 ≤ β ≤ φ+d and 0 ≤ d ≤ 1− 2φ. The leverage parameter γ ∈ (−1, 1) measures

the impact of asymmetric behavior of the residuals on the conditional volatility. The

power parameter δ determines whether it is appropriate to model the variance (δ = 2),

the standard deviation (δ = 1), or any other, real-valued degree. Ding et al. (1993) show

that autocorrelation of |rt|δ of a time series rt with δ close to 1 is larger and longer than

other values for δ and thus indicate the behavior of long memory in volatility. Another

possibility to measure long memory is given by fractionally integrated models such as

4

Table 1: Overview of univariate conditional variance models.

Model Definition Asymmetry Long memory

APARCH(1,1) hδ/2t = ω + α(|εt−1| − γεt−1)δ + βh

δ/2t−1 Yes Indirectly

FIAPARCH(1,d,1) hδ/2t = ω + (1− βL− (1− φL)(1− L)d)(|εt| − γεt)δ + βh

δ/2t−1 Yes Yes

FIAPARCH. The long memory parameter d ∈ (0, 1) reflects the persistence of shocks,

which rises when d approaches zero (Davidson, 2004).3

2.2. Dynamic Correlation Modeling

Analogously to the univariate modeling, we compare the properties of Bitcoin and

Gold from a multivariate perspective in terms of their relationship to other markets and

assets. Let Rt be a k-dimensional vector of observations at time t, denoted as

Rt = µt + εt

where µt is a k-dimensional conditional mean structure. For the k-dimensional vector εt,

we assume εt|Ft−1 ∼ N (0,Ht). We model conditional heteroskedasticity with

εt = H1/2t ζt, (1)

where Ht denotes the (k × k)-sized conditional variance matrix and ζt refers to a k-

dimensional vector of standard normally distributed and i.i.d. random variables with zero

mean and E[ζt, ζ′t ] = Ik. In Eq. (1), the matrix process Ht can be determined in several

ways; directly, with the Vector Error Correction model as suggested in (Bollerslev et al.,

1988) or with the Baba-Engle-Kraft-Kroner (BEKK-GARCH) model (Engle & Kroner,

1995), which is outlined subsequently.4 We apply the diagonal BEKK model which is a

sufficient compromise between parameter dimensionality and sample size. In the BEKK

model, the conditional variance-covariance matrix Ht is defined as

Ht = C>C + A>1 εt−1ε>t−1A + G>Ht−1G (2)

=

[c11 0

c12 c22

][c11 c12

0 c22

]+ diag [a11, a22]>

[ε2

1,t−1 ε1,t−1ε2,t−1

ε1,t−1ε2,t−1 ε22,t−1

]diag [a11, a22]

+ diag [g11, g22]>Ht−1diag [g11, g22] ,

(3)

3Note that we use the approach of Klein & Walther (2017) to estimate the parameter of fractionallyintegrated variance models. We implement an ARCH(∞) truncation lag of 1 000.

4Furthermore, the conditional covariance matrix can be modeled by decomposing Ht into conditionalstandard deviations and correlations, presented in Engle (2002) and referred to as the Dynamic Condi-tional Correlation (DCC) model. As we observe relatively extreme returns and variances in cryptocur-rencies (in comparison to other conventional assets), our DCC implementations are prone to instabilitiesand convergence issues.

5

where A,G, and C are k×k parameter matrices and C is lower triangular. In the diagonal

version of the BEKK defined in Eq. (3), matrices A and G are diagonal matrices. Non-

negativity, or positive-definiteness of Ht is achieved with relatively weak conditions (Engle

& Kroner, 1995).

Focusing on the most important linkages to other markets, we calculate the pairwise

correlation between Gold and Bitcoin to other assets by setting k = 2 and apply the BEKK

to centralized residuals. In Section 4, correlation plots are smoothed with a Savitzky-

Golay filter (Savitzky & Golay, 1964).

2.3. Portfolio-Based Comparison

In order to further investigate the hedging capabilities of Bitcoin and Gold, we imple-

ment a ex-post portfolio-based comparison with three steps:

1. We calculate the time-varying weights wt of a two component minimum-variance

portfolio of an asset with an market index, e.g. Bitcoin with S&P 500. The weights

are optimized for each point in time of our sample by

minwt

w′

tHtwt

s.t. w′

t1k = 1

where the covariance matrix Ht is obtained from the BEKK framework in Sec-

tion 4.2.

2. We calculate the historical Value-of-Risk of the S&P 500 or MSCI World over the

whole sample period, denoted by VaRq, i.e. we take the empirical quantile q at 1%,

5%, and 10% of the returns rt of that index. To obtain the VaR0.01 of an index, for

example, we sort all returns of that index in an ascending order. If the index has T

returns, the Value-at-Risk is the dT × 0.01e-th return in that list. Henceforth, we

define all points in time

t∗ := {t|rt < VaRq}

as times where the index is in distress.

3. Finally, we evaluate the two-component portfolio, consisting of an index and Gold

or Bitcoin, with returns rPFt calculated from the dynamic weights wt from step 1.

With a special focus on distressed times t∗, we calculate the mean of the portfolio

returns during these times as

rD =1

|t∗|∑m∈t∗

rPFm

which can be seen as a kind of Expected Shortfall or Conditional Value-at-Risk.

6

07/2011 01/2012 01/2013 01/2014 01/2015 01/2016 01/2017 01/2018

Time

1000

1100

1200

1300

1400

1500

1600

1700

1800

1900

Price

s in

US

D p

er

oz

1

10

100

1000

10000

100000

Price

s in

US

D p

er

BT

C

Gold

Bitcoin

Figure 1: Prices of Gold and Bitcoin(in logarithmic scale) from July 2011 to December 2017, n = 1 695.

This approach allows us to scrutinize whether the asset can serve as (temporary)

hedge, i.e. hedge the equity index during times of turmoil and thus, lower the impact of

distressed times on the index on average.

3. Data and Preliminary Analysis

In our analysis, we include a data set of six times series: the cryptocurrency Bitcoin,

Gold and Silver prices in USD per oz, crude oil prices for the West Texas Intermediate

(WTI), the S&P 500 index, MSCI World and the MSCI Emerging Markets 50 index. The

time series cover the period from 2011-07-01 to 2017-12-31 and are synchronized.5 From

daily closing prices, we calculate returns as the natural logarithmic price differences,

rt = 100 × log(Pt/Pt−1). For conventional assets, we take the closing price or index

points of each trading day. Data for these assets are obtained from Datastream with

GMT timestamp. For Bitcoin, we retrieve the data from coindesk.com, also with GMT

timestamp, and process the price at the end of the day since cryptocurrencies are traded

continuously. Since Bitcoin is also traded on weekends, we only consider the prices during

the week to synchronize our dataset.6 Hence, we obtain 1 695 returns for each time series.

Prices for Gold and Bitcoin are plotted in Fig. 1. This figure puts the extreme price

increases of Bitcoin in 2017 into perspective of its price history. In order to visualize this

dramatic increase, prices are plotted on a logarithmic scale.

Tab. 2 summarizes the descriptive statistics and some first time series tests for the

six assets. Here, Bitcoin returns have the highest mean and standard deviation by far.

5We choose this time window to avoid most of the zero-return trading days in the beginning of Bitcointrading.

6Note that we also tested weekend price interpolation for the conventional assets. However, thisintroduced some bias to the model estimation by low or virtually zero returns over the weekend.

7

Table 2: Descriptive statistics for Bitcoin and financial daily return time series for Jul 2, 2011 to Dec 31,2017, n = 1 695 observations.

Bitcoin Gold Silver WTI S&P 500 MSCI World MSCI EM50

Mean 0.4037 −0.0079 −0.0407 −0.0267 0.0489 0.0363 0.0245Std. Dev. 5.7577 1.0499 1.7633 2.1369 0.8859 0.8069 0.9444Min. −44.3784 −9.5962 −12.9970 −10.7263 −6.8801 −5.2465 −5.1962Max. 49.9663 4.8387 7.5760 11.6213 4.6344 4.1164 3.9169Skewness −0.3406 −0.6239 −0.7120 0.0956 −0.5306 −0.6622 −0.2586Kurtosis 14.2658 10.1569 8.5263 6.3419 9.0885 8.8237 5.1929

Jarque Bera 8996.3259∗∗∗ 3727.4769∗∗∗ 2300.1225∗∗∗ 791.3631∗∗∗ 2697.5676∗∗∗ 2519.1784∗∗∗ 358.5180∗∗∗

Ljung Box (25) 60.4956∗∗∗ 37.9578∗∗ 38.4509∗∗ 38.9043∗∗ 68.2019∗∗∗ 66.5574∗∗∗ 109.0498∗∗∗

ARCH (25) 264.1286∗∗∗ 52.2211∗∗∗ 148.6979∗∗∗ 263.3804∗∗∗ 497.2257∗∗∗ 379.9198∗∗∗ 248.6839∗∗∗

ADF −40.0953∗∗∗ −41.9713∗∗∗ −42.6236∗∗∗ −45.2808∗∗∗ −43.5181∗∗∗ −36.4768∗∗∗ −33.6676∗∗∗

Note: Std. Dev. is the standard deviation, Min. and Max. are minimum and maximum of the time series. ARCH(25) is thetest for autoregressive conditional heteroskedasticity by Engle (1982) at the 25th lag. ADF is the Augmented Dickey-Fullertest for unit root.

Moreover, the three commodities, Gold, Silver, and WTI, have negative mean returns

between −0.04% and −0.008%. The three equity indices, S&P 500, MSCI World, and

MSCI EM50, have slightly positive mean returns. However, among the five conventional

assets, the WTI has the highest daily standard deviation with 2.14%. Bitcoin has a daily

standard deviation of 5.76% which is more than two times of the risk of WTI. Skewness,

Kurtosis, and subsequently the Jarque Bera test result in the conclusion that none of

the six time series are Normally distributed. The Ljung-Box and the ARCH test suggest

autocorrelation in the returns and their volatility. Lastly, the Augmented Dickey-Fuller

test rejects the hypothesis of a unit root in the data and we assume all time series to be

stationary.

We illustrate the descriptive statistics of Bitcoin and Gold in Fig. 2. The left panel

compares the return series of both assets and elucidates the differences in volatility level.

The right panel captures the high kurtosis of both series in a histogram. It is obvious that

Bitcoin’s tails are much more pronounced than the tails of the density of Gold returns.

Since literature suggests bubble behavior in prices of Bitcoin, we briefly analyze the

time series as outlined in Kruse et al. (2018) and estimate

log(Pt) = a+ b log(Pt−1) + εt

on a rolling 30-day window.7 Explosive behavior, indicated by b > 1 for the respective

30-day window, is detected on isolated occasions which are spread throughout the history

of the Bitcoin price series. We note that not only the price increase of November and

December 2017 is detected as explosive behavior but also other significant—yet smaller

scale—price jumps, e.g. during the beginning of 2013. These jumps might explain the

7For a discussion on bias-reduction with indirect inference of this estimation method for shorter rollingwindows, we refer to Kruse et al. (2018).

8

07/201101/2012 01/2013 01/2014 01/2015 01/2016 01/2017 01/2018

Time

-40

-30

-20

-10

0

10

20

30

40R

etur

ns

BitcoinGold

-20 -15 -10 -5 0 5 10 15 20

Returns

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

BitcoinGold

Figure 2: Plots of the daily return series of Gold and Bitcoin (left) and resulting histogram (right). Forreadability, the ordinate of the return plot is limited to [−40, 40] and the abscissa of the histogram islimited to [−20, 20]. A small number of returns of Bitcoin is exceeding these limits.

Table 3: Unconditional pairwise Pearson correlation matrix for the sample period Jul 2, 2011 to Dec 31,2017.

Bitcoin Gold Silver WTI S&P 500 MSCI World MSCI EM50

Bitcoin 1.0000 0.0459 0.0071 0.0158 0.0491 0.0457 0.0042Gold 1.0000 0.6877 0.0655 −0.0366 0.0591 0.0840Silver 1.0000 0.1954 0.0775 0.2028 0.2181WTI 1.0000 0.3481 0.3800 0.2705S&P 500 1.0000 0.9148 0.5041MSCI World 1.0000 0.6706MSCI EM50 1.0000

high volatility of Bitcoin as well as the high mean return.

As a first indication for the correlation between the six assets, Tab. 3 presents the

unconditional pairwise Pearson correlations. We find that Bitcoin has only a slightly

positive correlation with the other assets, from which Gold, S&P 500 and MSCI World

have the highest correlation with Bitcoin between 0.045 and 0.049. The MSCI EM50

index has the lowest correlation with Bitcoin. Interestingly, Gold has a slightly negative

correlation with S&P 500, but a positive correlation of 0.059 with MSCI World, which

is similar to the correlation of Bitcoin to MSCI World. However, since this is only the

average correlation over the whole sample, we derive the time-varying specifics from the

BEKK model in Section 4.2.

4. Results

4.1. Comparison of Conditional Variance Dynamics

In a first step, we analyze and compare the univariate volatility structure of the six

assets and apply the APARCH and FIAPARCH model. These ARCH models can depict

long memory and asymmetric effects. We estimate the parameters of an AR(1)-APARCH

and -FIAPARCH model with Student’s t distributed errors with ν degrees-of-freedom.

9

Table 4: Estimation results from APARCH model with n = 1 695 observations. Statistically significantparameters are indicated with asterisk *, **, *** for 10%, 5%, and 1% level of significance.

Bitcoin Gold Silver WTI S&P 500 MSCI World MSCI EM50

θ0 0.3521∗∗∗ −0.0051 −0.0009 −0.0150 0.0582∗∗∗ 0.0399∗∗∗ 0.0205θ1 −0.0364 −0.0164 −0.0607∗∗∗ −0.0696∗∗∗ −0.0670∗∗∗ 0.0812∗∗∗ 0.1777∗∗∗

ω 0.8581 0.0054 0.0362 0.0089∗∗ 0.0287∗∗∗ 0.0206∗∗ 0.0197∗∗

α 0.1694∗∗∗ 0.0135 0.0027 0.0336∗∗∗ 0.1094∗∗∗ 0.0932∗∗∗ 0.0250∗

β 0.8306∗∗∗ 0.9721∗∗∗ 0.9847∗∗∗ 0.9663∗∗∗ 0.8906∗∗∗ 0.9061∗∗∗ 0.9210∗∗∗

γ −0.1972∗∗∗ −0.0368 −0.1748∗ 0.9995∗∗∗ 0.9995∗∗∗ 0.9929∗∗∗ 0.9982∗∗∗

δ 2.3065∗∗∗ 2.7788∗∗ 4.0000∗∗∗ 1.1554∗∗∗ 0.9004∗∗∗ 0.9007∗∗∗ 2.0480∗∗∗

ν 2.7895∗∗∗ 4.0728∗∗∗ 3.0291∗∗∗ 7.4318∗∗∗ 4.9035∗∗∗ 5.6001∗∗∗ 12.9275∗∗∗

LL −4757.74 −2276.95 −3144.77 −3387.27 −1806.08 −1641.80 −2123.75BIC 9574.96 4613.39 6349.02 6834.02 3671.64 3343.09 4306.99

Jarque Bera 14297.2533∗∗∗ 18806.0706∗∗∗ 2814.1535∗∗∗ 151.5134∗∗∗ 465.5090∗∗∗ 1091.3124∗∗∗ 29.4072∗∗∗

Ljung Box (25) 55.2186∗∗∗ 36.6557∗ 38.1835∗∗ 18.0121 22.9455 24.1004 21.6948ARCH (25) 83.4615∗∗∗ 4.8511 42.1987∗∗ 21.5143 57.9416∗∗∗ 28.3117 26.3618

Analyzing the APARCH parameter in Tab. 4, we find Bitcoin to be more similar to

Gold and Silver than to WTI, S&P 500, MSCI World or MSCI EM50 at first glance.

Even though Gold does not have a significant leverage parameter γ, the two metals and

Bitcoin share the same sign. All three assets tend to have higher volatility, if the former

day’s residual was positive, which is known as an inverse leverage effect—a prominent

feature of Gold and Silver. This is in contradiction to the findings of Dyhrberg (2016),

who does not find a significant asymmetric effect. Our own calculations suggest that

this is due to the fact that we use the Student’s t-distribution to account for heavy tails

present in the distribution of returns.8 The sensitivity of results regarding the underlying

distribution is also discussed in Baur et al. (2017) whose findings are in line with those

presented above. Thus, we conjecture that using a better suited distribution reveals the

property of asymmetric effects for Bitcoin as shown by Catania & Grassi (2017). On the

contrary, the other four assets have a high positive γ indicating that volatility rises if last

day’s return is negative. Similarly, the power parameter δ for WTI, S&P 500, and MSCI

World is around 1, i.e. modeling the standard deviation appears better rather than the

variance. The estimated δ for Gold and Bitcoin on the other hand, is in a range between

2.3 and 2.78. The ARCH test reveals that APARCH seems unable to diminish the whole

autocorrelation structure in the volatility of Bitcoin.

The estimated parameters for the FIAPARCH are shown in Tab. 5. All time series

reveal some degree of long memory, which is most pronounced in Gold and Silver. The

Bitcoin does have a significant long memory effect, but is not as persistent as the two metal

commodities. We find an asymmetric effect for Bitcoin’s volatility which is comparable

to the ones of Gold and Silver as shown before for APARCH. Again, the power parameter

δ is of similar size for Bitcoin, Gold, and Silver. While it is between 2 and 3 for the

8When repeating the regression with normally distributed errors, we can confirm and replicate thefindings of Dyhrberg (2016).

10

Table 5: Estimation results from FIAPARCH model with n = 1 695 observations. Statistically significantparameters are indicated with asterisk *, **, *** for 10%, 5%, and 1% level of significance.

Bitcoin Gold Silver WTI S&P 500 MSCI World MSCI EM50

θ0 0.3105∗∗∗ −0.0041 −0.0008 −0.0091 0.0588∗∗∗ 0.0463∗∗∗ 0.0148θ1 −0.0012 −0.0176 −0.0564∗∗∗ −0.0680∗∗∗ −0.0632∗∗∗ 0.0793∗∗∗ 0.1778∗∗∗

ω 0.8022 0.0890 0.6805 0.3037∗∗∗ 0.1492∗∗ 0.0941 0.1112∗∗

α 0.1081 0.3198∗∗∗ 0.4418∗∗ 0.0000 0.1452∗∗ 0.0749 0.0981d 0.7838∗∗∗ 0.3046∗∗∗ 0.1163∗ 0.9855∗∗∗ 0.5390∗∗∗ 0.4580∗∗∗ 0.2836∗∗∗

β 0.6024∗∗ 0.6207∗∗∗ 0.5178∗∗ 0.9468∗∗∗ 0.5510∗∗∗ 0.4670∗∗∗ 0.3397∗

γ −0.1360∗∗ −0.1522 −0.1961∗ 0.4547∗∗∗ 0.9198∗∗∗ 0.7219∗∗∗ 0.5129∗∗∗

δ 2.6932∗∗∗ 2.1678∗∗∗ 3.0000∗∗∗ 1.4318∗∗∗ 1.1730∗∗∗ 1.2801∗∗∗ 1.6842∗∗∗

ν 2.8546∗∗∗ 3.9589∗∗∗ 3.0201∗∗∗ 7.2083∗∗∗ 5.3033∗∗∗ 6.1160∗∗∗ 12.0667∗∗∗

LL −4719.27 −2279.77 −3146.65 −3391.91 −1783.92 −1634.84 −2125.46BIC 9505.46 4626.46 6360.21 6850.73 3634.75 3336.61 4317.85

Jarque Bera 10471.5744∗∗∗ 9792.5785∗∗∗ 1204.6351∗∗∗ 199.7367∗∗∗ 560.8414∗∗∗ 1370.1336∗∗∗ 36.4122∗∗∗

Ljung Box (25) 66.7220∗∗∗ 36.8645∗ 35.6686∗ 17.2492 19.7676 21.1447 21.2311ARCH (25) 19.2901 5.7398 26.6395 18.5606 19.1779 11.6784 26.5850

cryptocurrency and the two metals, the other three assets have estimated parameters

between 1 and 1.5. The last parameter we want to focus on is the degree-of-freedom of

the underlying Student’s t distribution. Here, the small ν for Bitcoin, Gold, and Silver

(between 2.85 and 3.95) correspond to the high kurtosis of these assets shown in Tab. 2.

Moreover, we conclude that the FIAPARCH model seems appropriate for all seven assets,

since it is able to diminish the autocorrelation in the volatility, i.e. the ARCH test is not

rejected.

Finally, when comparing the BIC of APARCH and FIAPARCH as a model choice

criteria for our sample, we conclude that the three commodities—Gold, Silver, WTI—

and the emerging market index MSCI EM50 are better modeled with APARCH which

purely models asymmetry, while the remaining three assets—Bitcoin, S&P 500, and MSCI

World—have a better in-sample performance with FIAPARCH which additionally models

long memory.

Hence, from the perspective of the volatility structure, Bitcoin shows evidence that

it has a similar asymmetric response observed in Gold and Silver. The long memory

parameter indicates a different persistence or long memory, which might be due to the

short and volatile history of Bitcoin and cryptocurrencies in general.

4.2. Correlation of Bitcoin to Financial Markets

Based on the BEKK-GARCH framework, we estimate dynamic correlations between

the considered assets.9 We begin our analysis by reporting the findings for Gold. In view of

the flight-to-quality phenomenon (Hammoudeh et al., 2010) and safe haven categorization

(Baur & Lucey, 2010), time of market turmoil or distress is of particular interest. However,

our sample from July 2011 to December 2017 is characterized by relatively stable growth

9Parameter estimates for all BEKK combinations in this section are available upon request.

11

rates of financial markets, proxied in this study by the S&P 500 and the MSCI World

index. For the S&P 500, we identify only two short periods of market distress and

significant, shock-like declines during the stock sell-off in August 2015 and again late

2015/early 2016 which is connected to crude oil prices dropping below $30 per barrel.

Fig. 3 plots the dynamic correlations of Gold and S&P 500 returns. The shaded area

in blue marks significant market downturns. Several conclusions can be drawn from the

estimates. Most importantly, we see that the correlations drop to negative values during

market distress. In the first downturn, we observe a change from positive to negative

correlation and a turnaround in the subsequent market recovery. The drop in correlation

during the second market distress situation (end of 2015, beginning of 2016) is quite

significant as (non-smoothed) correlations drop from 0.3 to −0.4 within a few trading

days. This behavior is expected if the safe haven property holds true and is clearly in

favor of the flight-to-quality hypothesis. Given that our sample does not include a major

and prolonged market downturn (such as the Financial Crisis), these two smaller shocks

suggest that the the flight-to-quality is still intact. More surprisingly, the plot might

also reveal that Gold has not been a safe haven according to the definition of Baur &

Lucey (2010) relative to the S&P 500 in recent years. From 2014 on, we mainly observe

negative correlations in times of stable upward movements of the S&P 500. According

to the definitions of Baur & Lucey (2010), Gold then acts more like a hedge than a safe

haven in recent years.

For the MSCI World index, the picture is somewhat different; we observe the same

turn to negative correlations in times of market downturn. However, the correlation is

mainly positive, in particular between 2012 and 2014, when Gold was very volatile.10

Turning to time-varying correlations of Bitcoin and the S&P 500, plotted in Fig. 4, we

firstly observe that the BEKK-GARCH correlation is extremely volatile and alternating

between positive and negative values. A possible reason for this jump-like behavior might

be the different absolute level of variance of the time series, where Bitcoin is characterized

by very high and erratic returns, see Fig. 2. In the beginning of our sample, approximately

until the end of 2013, the correlations are unstable and the alternating character becomes

apparent when comparing the smoothing algorithm to actual correlation values. The

erratic path makes an application infeasible during these times. This could be caused by

the BEKK model itself or by the low prices of Bitcoin and a higher-than-usual volatility

cluster which can also be observed in Fig. 2. However, we also note that correlations

become slightly less alternating and erratic from 2014 on. Henceforth, the correlation is

negative in general with a few positive spikes in the smoothed path. Non-smoothed values

peak up to 0.5 and get as low as −0.4.

What makes the Bitcoin – S&P 500 correlation fundamentally different from the cor-

10Additional plots are available in the Appendix.

12

07/2011 01/2012 01/2013 01/2014 01/2015 01/2016 01/2017 01/2018

Time

-0.4

-0.3

-0.2

-0.1

0

0.1

0.2

0.3

0.4

0.5

0.6

BE

KK

Cor

rela

tion ;

Gold - S&P500

BEKK ;SG-smoothed ;

Figure 3: Dynamic correlations of Gold and S&P 500 returns obtained with the BEKK-GARCH betweenJuly 2, 2011 and December 31, 2017, n = 1 695. Unfiltered correlations are plotted in gray, Savitzky-Golay-smoothing is plotted in red. Times of market distress is highlighted in blue.

relations of Gold and the index is the behavior during market distress. Interestingly,

correlations are steeply increasing from negative to a positive relationship while the index

is in a downward movement. This indicates that Bitcoin follows the downturn, which

is observable in the raw as well as smoothed correlations in Fig. 4. The same behav-

ior holds for the Bitcoin – MSCI World correlations. While Gold prices increase in the

flight-to-quality, Bitcoin prices are decreasing with the markets.

To further highlight the differences, Fig. 5 visualizes the smoothed correlations of

Gold and Bitcoin with the S&P 500. Interestingly, the movements in correlations appear

to be mirrored from 2015 on, while being negative on average. This falls into the time

where Bitcoin is becoming more popular and price increases begin to accelerate. From

the joint plot, it becomes clear that Bitcoin, viewed as an asset, behaves differently than

Gold. Comparing the correlations of Gold and Bitcoin with the MSCI World, plotted

in the Appendix, the mirrored movements are more emphasized and span over different

signs. It appears that as soon as correlations of Gold and the MSCI World turn positive,

Bitcoin correlations become negative and vice versa. Similar behavior is observable for

the linkages to Silver, while the Gold – Silver correlation is naturally very high, and other

commodities such as the WTI.11 For correlations of the MSCI EM50 index, we observe

similar correlations of Gold and Bitcoin with this index from the second half of 2015 to

approximately the third quarter of 2016. During these times, Bitcoin quickly turned to

negative correlations during distress in the markets of this index. As we have an intact

flight-to-quality for Gold, we hence observe a temporal hedging situation. Before 2015

and from 2016:Q4 we observe the same mirroring of correlations as for the S&P 500 and

11The pots of the correlation of Bitcoin with MSCI World, MSCI EM50, and WTI can be found in theAppendix.

13

07/2011 01/2012 01/2013 01/2014 01/2015 01/2016 01/2017 01/2018

Time

-0.4

-0.3

-0.2

-0.1

0

0.1

0.2

0.3

0.4

0.5

0.6

BE

KK

Cor

rela

tion ;

Bitcoin - S&P500

BEKK ;SG-smoothed ;

Figure 4: Dynamic correlations of Bitcoin and S&P 500 returns obtained with the BEKK-GARCHbetween July 2, 2011 and December 31, 2017, n = 1 695. Unfiltered correlations are plotted in gray,Savitzky-Golay-smoothing is plotted in red. Times of market distress is highlighted in blue.

the MSCI World index.

Rounding up the analysis, Fig. 6 plots the dynamic correlations of Bitcoin and Gold

prices. Interestingly, we observe a coupling effect in 2013 during the price decay of Gold.

In 2014, we see some decoupling in the first half and positive correlations in the second

half. End of 2013, Bitcoin spikes for the first time at around 1 150 US Dollar and is in a

decline during 2014, accompanied by some minor price recoveries. For Gold, prices recover

from 1 200 USD/oz to almost 1 400 USD/oz in the first quarter while being generally

decreasing throughout the year. This volatile period of Gold prices continues in 2015 while

Bitcoin has relatively small price movements (relative to its own price history and volatility

clustering). Hence we observe lower correlations which are somewhat stable. Focusing on

2017 with the Bitcoin price explosion beginning in the second quarter, Bitcoin and Gold

are uncorrelated on average in the second and third quarter. In the fourth quarter, they

couple again as we first observe strong price increases and shortly after price drops in

Bitcoin that correlate with upward Gold price movements. We conclude that Bitcoin and

Gold feature no stable correlation. Their return relationship is characterized by positive

and negative spikes with no general tendency. One would assume this correlation to

be positive and stable if Bitcoin is believed to be similar to Gold. If we compare Gold

with Silver, another precious metal utilized as investment, we observe high and stable

correlations.

Concluding our correlation analysis, we find Gold to be a hedge rather than a safe

haven in recent years. Bitcoin, on the other hand, behaves completely different, especially

from 2015 on. The cryptocurrency couples with markets during bearish environments,

with correlations rapidly turning to positive values in these times. This holds true for

both the S&P 500 and the MSCI World index. During the last two years, we also observe

14

07/2011 01/2012 01/2013 01/2014 01/2015 01/2016 01/2017 01/2018

Time

-0.4

-0.3

-0.2

-0.1

0

0.1

0.2

0.3

0.4

0.5

0.6

BE

KK

Cor

rela

tion ;

Gold - S&P500 and Bitcoin - S&P500 Correlation

Gold - S&P500BTC - S&P500

Figure 5: Smoothed correlations of Bitcoin and Gold returns with S&P 500 returns obtained with theBEKK-GARCH between July 2, 2011 and December 31, 2017, n = 1 695. Times of market distress ishighlighted in blue.

inverse movements of correlations of Gold and Bitcoin with these two indices. While

correlations increase for Gold, Bitcoin correlations decrease to the same market and vice

versa. This is a clear indication that Bitcoin and Gold have different connectedness to

markets. However, we observe a temporal similarity of Bitcoin and Gold correlations

with the MSCI EM50 index from mid-2015 to mid-2016. This effect is short-lived and

dissipates after to mirrored correlations as in other markets. This effect could be due to

the high acceptance of cryptocurrencies in the countries comprising the emerging markets

index e.g. China. Applying the categorizations of Baur & Lucey (2010), we find Bitcoin

to fit neither of them as we observe phases of positive and negative correlations, quickly

alternating, as well as a positive correlation in market turmoil which argues against safe

haven and hedging properties. This supports findings of Bouri et al. (2017b) based on

data up to December 2015, that identify no safe haven or hedging properties against the

assets in the present study.

4.3. Portfolio-based Test of Hedging Property

In this part, we analyze the hedging property of Bitcoin and Gold. The portfolios under

examination are: Bitcoin/S&P 500, Gold/S&P 500, Bitcoin/MSCI World, Gold/MSCI

World, Bitcoin/MSCI EM50, and Gold/MSCI EM50. In the following, we proceed as

outlined in Section 2.3.

Firstly, we calculate the time-varying portfolio weights of the two component portfo-

lios. The weights for the four portfolios are presented in Panel A of Tab. 6. Additionally,

we illustrate the weights for the two portfolios with the S&P 500 index in Fig. 7. It can

be seen that the Gold proportion has a high variation within the portfolios with the two

equity indices. This results in extreme cases where Gold tends to have almost 90% of the

portfolio or is even short sold. The average Gold weight is 36.98% for S&P 500, 30.88%

15

07/2011 01/2012 01/2013 01/2014 01/2015 01/2016 01/2017 01/2018Time

-0.4

-0.3

-0.2

-0.1

0

0.1

0.2

0.3

0.4

0.5

0.6

BE

KK

Cor

rela

tion

Bitcoin - Gold

BEKK SG-smoothed

Figure 6: Dynamic correlations of Bitcoin and Gold returns obtained with the BEKK-GARCH betweenJuly 2, 2011 and December 31, 2017, n = 1 695. Unfiltered correlations are plotted in gray, Savitzky-Golay-smoothing is plotted in red.

for MSCI World, and 42.70% for MSCI EM50, respectively. Confirming the results of

Guesmi et al. (2018), we find that Bitcoin only has a small proportion of a joint portfo-

lio with the three stock indices and averages between 2.92% and 4.78%, while having a

standard deviation of about the same size. Obviously, this behavior stems from the high

volatility of Bitcoin returns. For some periods for Gold and for Bitcoin, we obtain short

positions, which can be seen by the minimum weights and in Fig. 7. The possibility to

short Bitcoin is only a very recent development, gaining popularity with the introduction

of Bitcoin futures on the CBOE and CME in December 2017. Until then, short-selling in

the portfolio composition should be restricted. However, this also eliminates previously

mentioned hedging possibilities by shorting Bitcoin investments. Given a functioning

and liquid futures exchange, negative weights in Bitcoin investments should be seen as a

possible investment opportunity in the future in view of portfolio variance reduction.

Secondly, the historical Value-at-Risk measures are determined and given in Panel B

of Tab. 6. We find that the values for the S&P 500 are slightly lower than for the MSCI

World. The MSCI EM50 index has the lowest 5% and 10% Value-at-Risk measures.

However, the extreme case of 1% is still above the Value-at-Risk of the S&P 500.

Lastly, we use the calculated Value-at-Risk values to determine times of distress for

the two equity markets and evaluate how a portfolio including Bitcoin or Gold reacts in

these times of turmoil. The Panel C in Tab. 6 reports the average return, the volatility,

and the average return of the portfolios for the times when the index is below the Value-

at-Risk of that index. More specifically, we report a 100% investment in the portfolio

compared with the minimum-variance portfolios. From the results in Panel C, we can

draw the following two conclusions:

1) Over the whole sample, the minimum-variance combination of the indices with

16

Table 6: Statistics for the minimum-variance portfolios for Bitcoin between Jul 2, 2011 and Dec 31, 2017.

Panel A: Descriptive statistics of the portfolio weights

S&P 500 MSCI World MSCI EM50Bitcoin Gold Bitcoin Gold Bitcoin Gold

Mean 0.0369 0.3698 0.0292 0.3088 0.0478 0.4270Std. Dev. 0.0403 0.1368 0.0334 0.1451 0.0439 0.1157Min. −0.0850 −0.0554 −0.0487 −0.0429 −0.0486 −0.1208Max. 0.2203 0.8896 0.1585 0.7883 0.1929 0.7363

Panel B: Value-at-Risk measures

S&P 500 MSCI World MSCI EM50

VaR0.01 −2.5438 −2.3171 −2.5096VaR0.05 −1.4159 −1.2454 −1.5707VaR0.10 −0.9019 −0.8031 −1.1353

Panel C: Hedging Properties

S&P 500 Bitcoin Gold MSCI World Bitcoin Gold MSCI EM50 Bitcoin Gold

Return 0.0489 0.0547 0.0251 0.0363 0.0399 0.0239 0.0245 0.0370 0.0118Volatility 0.8859 0.9044 0.6483 0.8069 0.8174 0.6434 0.9444 0.9361 0.7265Return|VaR0.01 −3.5755 −3.6577 −1.0105 −3.4736 −3.4840 −1.1856 −3.2653 −3.0775 −1.0866Return|VaR0.05 −2.1881 −2.2189 −1.0195 −2.0255 −2.0267 −1.1387 −2.1592 −2.0828 −1.1383Return|VaR0.10 −1.6597 −1.6619 −0.8616 −1.5264 −1.5147 −0.8835 −1.7452 −1.6628 −1.0018

Bitcoin slightly increases the average return as well as the volatility of the re-

turns. Only for the combination with the MSCI EM50 index, the volatility of

the minimum-variance portfolio remains almost unchanged. On the contrary, the

minimum-variance portfolios of Gold and the indices decrease both, average return

and volatility.

2) During the time where the equity indices are under distress, i.e. where the return

of the index is smaller than the Value-at-Risk calculated in Panel B of Tab. 6,

Bitcoin does not provide any kind of hedge. For the S&P 500 and MSCI World,

the minimum-variance combinations of the equity indices with Bitcoin lead to even

lower average returns compared to a sole investment in the index, except for the 10%

Value-at-Risk of MSCI World. Only the combination with the MSCI EM50 index

leads to a slightly higher returns. On the other hand, Gold provides that hedge,

which confirms the findings of the previous section. Gold may not fully protect

from market drawdowns, but the hedging properties are clearly pronounced. The

minimum-variance portfolios of Gold and the indices have a higher average return

during times under distress.

Hence, we conclude from our ex-post portfolio exercise that in contrast to Gold, Bitcoin

does not provide a hedge for equity investments.

4.4. Robustness Check with Cryptocurrency Index

In order to test whether our results hold for other cryptocurrencies, we repeat our anal-

ysis with the broad, market-weighted cryptocurrency index CRIX (Trimborn & Hardle,

2016). We retrieve the index from July 31, 2014 to December 31, 2017 from the website

17

07/2011 01/2012 01/2013 01/2014 01/2015 01/2016 01/2017 01/2018

Time

-0.1

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

Wei

ghts

Minimum Variance Composition

Gold weightsBTC weights

Figure 7: Time-varying weights of minimum-variance portfolios for Gold–S&P 500 and Bitcoin–S&P 500based on BEKK correlations.

crix.hu-berlin.de. We synchronize the index series in the same fashion (e.g. excluding

weekends) as before for the Bitcoin prices. Hence, we calculate 891 returns beginning

August 1, 2014.

The descriptive statistics and the correlation matrix for CRIX are given in the Ap-

pendix in Tab. A.8 and Tab. A.9. The former are similar to those we present for Bitcoin.

For the unconditional correlation, we find slight changes in comparison with Tab. 3.

Firstly, the pairwise correlation of Gold and the two MSCI indices is negative. Secondly,

CRIX has a small negative correlation with the S&P500 and MSCI EM50 returns. How-

ever, this only stems from the shorter sample period and might not be related to the

choice of CRIX.

As Bitcoin has the largest share in CRIX, we obtain similar results for CRIX. We find

significant asymmetries with APARCH in Tab. A.10. Interestingly, the estimations for

FIAPARCH presented in Tab. A.11 might reveal spurious long memory (Walther et al.,

2017). However, it is out of the scope of this paper to further address this issue.

The correlation of CRIX and global markets is almost identical than the reported

relationship of Bitcoin. Despite a varying weight of Bitcoin in CRIX, with other cryp-

tocurrencies such as Ethereum taking significant shares in the recent months, it appears

that during times of market distress cryptocurrencies follow market movements. From

the results for CRIX, we also conjecture that cryptocurrencies are highly correlated with

each other, in particular when facing a general downward movement of equity markets.

Finally, we repeat our portfolio analysis with the cryptocurrency index CRIX. The

results are given in Tab. 7. The corresponding portfolio weights for CRIX and the S&P 500

are plotted in Fig. A.8. In contrast to Bitcoin, the CRIX seems to have marginal hedging

effects for the three equity indices. During times of distress, portfolio returns are slightly

better. However, comparing it with the hedging performance of Gold during that period,

18

CRIX still fails to be considered an effective hedge. However, we highlight that in contrast

to Bitcoin, CRIX offers some potential in view of portfolio volatility reduction. Comparing

Panel C of Tab. 6 and Tab. 7, we find that CRIX reduces the portfolio’s volatility and

slightly increases the overall return. Bitcoin also increases the portfolio return at the cost

of an increased volatility. Although, it is important to note that Tab. 6 and Tab. 7 are

based on a different time frame.

Table 7: Statistics for the minimum-variance portfolios for CRIX between Aug 1, 2014 and Dec 31, 2017.

Panel A: Descriptive statistics of the portfolio weights

S&P 500 MSCI World MSCI EM50CRIX Gold CRIX Gold CRIX Gold

Mean 0.0530 0.4039 0.0421 0.3425 0.0724 0.5053Std. Dev. 0.0509 0.1275 0.0466 0.1332 0.0615 0.0956Min. −0.0951 0.1287 −0.1175 0.1070 −0.1012 0.3078Max. 0.2594 0.8131 0.2421 0.7813 0.2455 0.7847

Panel B: Value-at-Risk measures

S&P 500 MSCI World MSCI EM50

VaR0.01 −2.1599 −2.0520 −2.6449VaR0.05 −1.3051 −1.0624 −1.4866VaR0.10 −0.8162 −0.7026 −1.0590

Panel C: Hedging Properties

S&P 500 CRIX Gold MSCI World CRIX Gold MSCI EM50 CRIX Gold

Return 0.0446 0.0507 0.0221 0.0324 0.0368 0.0203 0.0363 0.0530 0.0163Volatility 0.7637 0.7601 0.5053 0.6835 0.6810 0.4920 0.9009 0.8694 0.5914Return|VaR0.01 −2.8932 −3.0436 −0.8847 −2.7980 −2.6341 −0.8606 −3.1531 −2.7419 −0.4558Return|VaR0.05 −1.8702 −1.8388 −0.6936 −1.6892 −1.6474 −0.6314 −2.0633 −1.8402 −0.7102Return|VaR0.10 −1.4450 −1.4140 −0.6137 −1.2906 −1.2344 −0.6150 −1.6555 −1.4821 −0.6961

5. Conclusion

Based on recent data up to December 2017, we revisit the conditional volatility mod-

eling of Bitcoin returns and update the results of previous applications of GARCH-class

models on cryptocurrency prices. Our results indicate that the proper choice of the un-

derlying distribution in these models is very important regarding the interpretation of

identified stylized facts. In contrast to other studies, we find FIAPARCH to be the best

fitting model in terms of log-likelihood and information criteria. This implies that Bitcoin

returns have an asymmetric response to market shocks, which is of the same direction

than precious metals. Price increases lead to an increase in volatility. Considering the

extreme price increases observed for Bitcoin, this finding is not surprising. The high per-

sistence of variance shocks, however, is the more dominating property and indicates that

volatility declines slowly after an increase. In Bitcoin markets, we observed a tremendous

rally in 2017 ending up with a major setback. These periods of elevated volatility lead to

a higher persistence.

From a perspective of market linkages, our correlation modeling provides evidence that

Bitcoin behaves completely different from Gold, in particular in market distress. While

19

the flight-to-quality property of Gold is confirmed, Bitcoin shows a positive coupling effect

and declines when markets are declining in shock-like situations. This is affirmed in a

portfolio application which suggests that Bitcoin is no hedge against equity investments.

However, our sample size is limited and we only observe a very small number of these

downturns. Given the relatively young markets, this statement should be tested again

in matured cryptocurrency markets. For now, Bitcoin as an asset does not resemble any

other conventional asset from an econometric perspective. Our results hold for other

cryptocurrencies and a short time period.

We believe that cryptocurrencies will remain highly volatile and continue to exhibit

strong movements to both directions as future development stays highly unclear. Signif-

icant price movement of cryptocurrencies are dependent on several factors. First, cryp-

tocurrencies will continue to experience large drops in price as investors will continue to

take profit at the peak of price movement. Second, regulatory decisions will have a strong

impact on investors’ behavior. Currently, regulatory agencies are still weighing on a legal

frame for cryptocurrencies. Third, in the face of recurring cyber attacks, such as on Mt-

Gox, Instawallet, or Bithump, the cryptocurrency ecosystem will have to strengthen its

security standards to become accepted by traditional investors.

20

References

Baillie, R. T. (1996). Long memory processes and fractional integration in econometrics.Journal of Econometrics , 73 , 5–59. doi:10.1016/0304-4076(95)01732-1.

Bariviera, A. F., Basgall, M. J., Hasperue, W., & Naiouf, M. (2017). Some stylized facts ofthe Bitcoin market. Physica A: Statistical Mechanics and its Applications , 484 , 82–90.doi:10.1016/j.physa.2017.04.159.

Baur, D. G., Dimpfl, T., & Kuck, K. (2017). Bitcoin, gold and the US dollar – A replicationand extension. Finance Research Letters, forthcoming , . doi:10.1016/j.frl.2017.10.012.

Baur, D. G., & Lucey, B. M. (2010). Is Gold a Hedge or a Safe Haven? An Analysis ofStocks, Bonds and Gold. Financial Review , 45 , 217–229. doi:10.1111/j.1540-6288.2010.00244.x.

Baur, D. G., & McDermott, T. K. (2010). Is gold a safe haven? International evidence.Journal of Banking & Finance, 34 , 1886–1898. doi:10.1016/j.jbankfin.2009.12.008.

Baur, D. G., & McDermott, T. K. (2016). Why is gold a safe haven? Journal of Behavioraland Experimental Finance, 10 , 63–71. doi:10.1016/j.jbef.2016.03.002.

Black, F. (1976). Studies of Stock Price Volatility Changes. In Proceedings of the Businessand Economics Section of the American Statistical Association (p. 177–181).

Blau, B. M. (2017). Price dynamics and speculative trading in bitcoin. Research inInternational Business and Finance, 41 , 493–499. doi:10.1016/j.ribaf.2017.05.010.

Bollerslev, T. (1986). Generalized autoregressive conditional heteroskedasticity. Journalof Econometrics , 31 , 307–327. doi:10.1016/0304-4076(86)90063-1.

Bollerslev, T., Engle, R. F., & Wooldridge, J. M. (1988). A Capital Asset Pricing Modelwith Time-varying Covariances. Journal of Political Economy , 96 , 116–131.

Bollerslev, T., & Mikkelsen, H. O. (1996). Modeling and pricing long memory in stockmarket volatility. Journal of Econometrics , 73 , 151–184. doi:10.1016/0304-4076(95)01736-4.

Bouri, E., Gupta, R., Tiwari, A. K., & Roubaud, D. (2017a). Does Bitcoin hedge globaluncertainty? Evidence from wavelet-based quantile-in-quantile regressions. FinanceResearch Letters , 23 , 87–95. doi:10.1016/j.frl.2017.02.009.

Bouri, E., Molnar, P., Azzi, G., Roubaud, D., & Hagfors, L. I. (2017b). On the hedge andsafe haven properties of Bitcoin: Is it really more than a diversifier? Finance ResearchLetters , 20 , 192–198. doi:10.1016/j.frl.2016.09.025.

Brandvold, M., Molnar, P., Vagstad, K., & Andreas Valstad, O. C. (2015). Price discoveryon Bitcoin exchanges. Journal of International Financial Markets, Institutions andMoney , 36 , 18–35. doi:10.1016/j.intfin.2015.02.010.

21

Campbell, J. Y., & Hentschel, L. (1992). No news is good news. An asymmetric modelof changing volatility in stock returns. Journal of Financial Economics , 31 , 281–318.doi:10.1016/0304-405X(92)90037-X.

Capie, F., Mills, T. C., & Wood, G. (2005). Gold as a hedge against the dollar. Journal ofInternational Financial Markets, Institutions and Money , 15 , 343–352. doi:10.1016/j.intfin.2004.07.002.

Catania, L., & Grassi, S. (2017). Modelling Crypto-Currencies Financial Time-Series.URL: https://www.ssrn.com/abstract=3028486. doi:10.2139/ssrn.3028486.

Cheah, E. T., & Fry, J. (2015). Speculative bubbles in Bitcoin markets? An empiricalinvestigation into the fundamental value of Bitcoin. Economics Letters , 130 , 32–36.doi:10.1016/j.econlet.2015.02.029.

Cheung, A. W.-K., Roca, E., & Su, J.-J. (2015). Crypto-currency bubbles: an applica-tion of the Phillips–Shi–Yu (2013) methodology on Mt. Gox bitcoin prices. AppliedEconomics , 47 , 2348–2358. doi:10.1080/00036846.2015.1005827.

Chkili, W., Hammoudeh, S., & Nguyen, D. K. (2014). Volatility forecasting and riskmanagement for commodity markets in the presence of asymmetry and long memory.Energy Economics , 41 , 1–18. doi:10.1016/j.eneco.2013.10.011.

Christie, A. A. (1982). The stochastic behavior of common stock variances. Value, leverageand interest rate effects. Journal of Financial Economics , 10 , 407–432. doi:10.1016/0304-405X(82)90018-6.

Chu, J., Chan, S., Nadarajah, S., & Osterrieder, J. (2017). GARCH Modelling ofCryptocurrencies. Journal of Risk and Financial Management , 10 , 17. doi:10.3390/jrfm10040017.

Chuen, D. L. K. (Ed.) (2015). Handbook of Digital Currency: Bitcoin, Innovation, Fi-nancial Instruments, and Big Data. London: Academic Press.

Ciaian, P., Rajcaniova, M., & Kancs, D. (2017). Virtual relationships: Short- and long-run evidence from BitCoin and altcoin markets. Journal of International FinancialMarkets, Institutions and Money , 6 , 467–486. doi:10.1016/j.intfin.2017.11.001.

Conrad, C., Custovic, A., & Ghysels, E. (2018). Long-and Short-Term CryptocurrencyVolatility Components: A GARCH-MIDAS Analysis. Journal of Risk and FinancialManagement , 11 , 1–12. doi:10.3390/jrfm11020023.

Corbet, S., Meegan, A., Larkin, C., Lucey, B., & Yarovaya, L. (2018). Exploring the Dy-namic Relationships between Cryptocurrencies and Other Financial Assets. EconomicsLetters , 165 , 28–34. doi:10.1016/j.econlet.2018.01.004.

Davidson, J. (2004). Moment and Memory Properties of Linear Conditional Heteroscedas-ticity Models, and a New Model. Journal of Business & Economic Statistics , 22 , 16–29.doi:10.1198/073500103288619359.

Demir, E., Gozgor, G., Lau, C. K. M., & Vigne, S. A. (2018). Does economic policyuncertainty predict the Bitcoin returns? An empirical investigation. Finance ResearchLetters , . doi:10.1016/j.frl.2018.01.005.

22

Ding, Z., Granger, C. W. J., & Engle, R. F. (1993). A long memory property of stockmarket returns and a new model. Journal of Empirical Finance, 1 , 83–106. doi:10.1016/0927-5398(93)90006-D.

Dwyer, G. P. (2015). The economics of Bitcoin and similar private digital currencies.Journal of Financial Stability , 17 , 81–91. doi:10.1016/j.jfs.2014.11.006.

Dyhrberg, A. H. (2016). Bitcoin, gold and the dollar – A GARCH volatility analysis.Finance Research Letters , 16 , 85–92. doi:10.1016/j.frl.2015.10.008.

Elendner, H., Trimborn, S., Ong, B., & Lee, T. M. (2018). The Cross-Section of Crypto-Currencies as Financial Assets, Volume 1. In Handbook of Blockchain, Digital Finance,and Inclusion (pp. 145–173). Elsevier volume 649. doi:10.1016/B978-0-12-810441-5.00007-5.

Engle, R. F. (1982). Autoregressive Conditional Heteroscedasticity with Estimates ofthe Variance of United Kingdom Inflation. Econometrica, 50 , 987–1007. doi:10.2307/1912773.

Engle, R. F. (2002). Dynamic Conditional Correlation. A Simple Class of MultivariateGeneralized Autoregressive Conditional Heteroskedasticity Models , 20 , 339–350. doi:10.1198/073500102288618487.

Engle, R. F., & Kroner, K. F. . (1995). Multivariate Simultaneous Generalized Arch.Econometric Theory , 11 , 122–150. doi:10.1017/S0266466600009063.

Engle, R. F., & Ng, V. K. (1993). Measuring and Testing the Impact of News on Volatility.The Journal of Finance, 48 , 1749–1778. doi:10.2307/2329066.

Guesmi, K., Saadi, S., Abid, I., & Ftiti, Z. (2018). Portfolio diversification with virtualcurrency: Evidence from bitcoin. International Review of Financial Analysis , . doi:10.1016/j.irfa.2018.03.004.

Hammoudeh, S. M., Yuan, Y., McAleer, M., & Thompson, M. A. (2010). Precious metals-exchange rate volatility transmissions and hedging strategies. International Review ofEconomics & Finance, 19 , 633–647. doi:10.1016/j.iref.2010.02.003.

Hardle, W. K., Harvey, C. R., & Reule, R. C. G. (2018). Understanding Cryptocurrencies?

Katsiampa, P. (2017). Volatility estimation for Bitcoin: A comparison of GARCH models.Economics Letters , 158 , 3–6. doi:10.1016/j.econlet.2017.06.023.

Klein, T. (2017). Dynamic Correlation of Precious Metals and Flight-to-Quality in De-veloped Markets. Finance Research Letters , 23 , 283–290. doi:10.1016/j.frl.2017.05.002.

Klein, T., & Walther, T. (2017). Fast fractional differencing in modeling long memory ofconditional variance for high-frequency data. Finance Research Letters , 22C , 274–279.doi:10.1016/j.frl.2016.12.020.

Kruse, R., Kaufmann, H., & Wegener, C. (2018). Bias-corrected estimation for speculativebubbles in stock prices. Economic Modelling , forthc., 1–11. doi:10.1016/j.econmod.2018.04.014.

23

Nakamoto, S. (2008). Bitcoin: A Peer-to-Peer Electronic Cash System. URL: https://bitcoin.org/bitcoin.pdf.

Savitzky, A., & Golay, M. J. E. (1964). Smoothing and Differentiation of Data by Sim-plified Least Squares Procedures. Analytical Chemistry , 36 , 1627–1639. doi:10.1021/ac60214a047.

Trimborn, S., & Hardle, W. K. (2016). CRIX an Index for blockchain based Currencies.URL: http://crix.hu-berlin.de/data/CRIXpaper.pdf.

Tse, Y. K. (1998). The conditional heteroscedasticity of the yen-dollar exchange rate.Journal of Applied Econometrics , 13 , 49–55. doi:10.1002/(SICI)1099-1255(199801/02)13:1<49::AID-JAE459>3.0.CO;2-O.

Walther, T., Klein, T., Pham Thu, H., & Piontek, K. (2017). True or spurious longmemory in European Non-EMU currencies. Research in International Business andFinance, 40C , 217–230. doi:10.1016/j.ribaf.2017.01.003.

Zakoian, J. M. (1994). Threshold heteroskedastic models. Journal of Economic Dynamicsand Control , 18 , 931–955. doi:10.1016/0165-1889(94)90039-6.

24

Appendix A. Additional Figures and Tables for the Analysis of CRIX

To be made available in the supplementary material.

Table A.8: Descriptive statistics for CRIX and financial daily return time series for Aug 1, 2014 to Dec31, 2017, n = 891 observations.

CRIX Gold Silver WTI S&P 500 MSCI World MSCI EM50

Mean 0.4346 0.0010 −0.0210 −0.0545 0.0446 0.0324 0.0363Std. Dev. 4.2001 0.8682 1.5136 2.5811 0.7637 0.6835 0.9009Min. −23.8364 −3.1910 −7.7286 −10.7263 −4.0182 −5.0273 −4.4552Max. 19.8541 4.1964 5.4035 11.6213 3.8369 2.5681 3.4746Skewness −0.3455 0.2366 −0.3274 0.1405 −0.3599 −0.8015 −0.3213Kurtosis 9.0743 5.2371 5.9894 5.0188 6.4371 8.6975 4.5726

Jarque Bera 1387.5104∗∗∗ 194.1075∗∗∗ 347.6768∗∗∗ 154.2378∗∗∗ 457.8316∗∗∗ 1300.5326∗∗∗ 107.1454∗∗∗

Ljung Box (25) 40.2952∗∗ 49.3313∗∗∗ 49.8255∗∗∗ 28.6566 25.4508 53.6869∗∗∗ 46.8096∗∗∗

ARCH (25) 94.8775∗∗∗ 64.4326∗∗∗ 40.2444∗∗ 114.1034∗∗∗ 169.5419∗∗∗ 100.7778∗∗∗ 111.2621∗∗∗

ADF −28.9508∗∗∗ −30.3975∗∗∗ −31.0810∗∗∗ −33.2704∗∗∗ −30.0854∗∗∗ −25.5438∗∗∗ −24.8281∗∗∗

Note: Std. Dev. is the standard deviation, Min. and Max. are minimum and maximum of the time series. ARCH(25) is thetest for autoregressive conditional heteroskedasticity by Engle (1982) at the 25th lag. ADF is the Augmented Dickey-Fullertest for unit root.

Table A.9: Unconditional pairwise Pearson correlation matrix for the sample period Aug 1, 2014 to Dec31, 2017.

CRIX Gold Silver WTI S&P 500 MSCI World MSCI EM50

CRIX 1.0000 0.0438 0.0044 0.0272 −0.0028 0.0046 −0.0139Gold 1.0000 0.6644 0.0150 −0.1612 −0.1107 −0.0922Silver 1.0000 0.1606 0.0092 0.1023 0.1014WTI 1.0000 0.3013 0.3296 0.2064S&P 500 1.0000 0.9087 0.4599MSCI World 1.0000 0.6235MSCI EM50 1.0000

25

Table A.10: Estimation results from APARCH model with n = 891 observations (Aug 1, 2014-Dec 31,2017). Statistically significant parameters are indicated with asterisk *, **, *** for 10%, 5%, and 1%level of significance.

CRIX Gold Silver WTI S&P 500 MSCI World MSCI EM50

θ0 0.4094∗∗∗ −0.0101 0.0182 −0.0863∗∗∗ 0.0544∗∗ 0.0332∗∗∗ 0.0247θ1 −0.0386∗∗∗ −0.0050 −0.0279 −0.1186∗∗∗ −0.0903∗∗ 0.0661∗∗ 0.2099∗∗∗

ω 0.0967∗∗∗ 0.0030 0.0000 0.0139∗∗∗ 0.0379∗∗∗ 0.0231∗∗ 0.0716∗

α 0.2004∗∗∗ 0.0277 0.0017 0.0240∗∗∗ 0.1358∗∗∗ 0.1057∗∗∗ 0.0986∗

β 0.7996∗∗∗ 0.9709∗∗∗ 0.9941∗∗∗ 0.9760∗∗∗ 0.8641∗∗∗ 0.8943∗∗∗ 0.7891∗∗∗

γ −0.1614 0.0796 −0.9995 0.9995∗∗∗ 0.9995∗∗∗ 0.9995∗∗∗ 0.5615∗∗

δ 0.4354∗∗∗ 1.9657∗∗∗ 2.7273∗∗∗ 0.3106∗∗∗ 0.7967∗∗∗ 0.8513∗∗∗ 1.8599∗∗∗

ν 3.0187∗∗∗ 4.2659∗∗∗ 2.8387∗∗∗ 10.6260∗∗∗ 4.6021∗∗∗ 5.5958∗∗∗ 15.4109∗∗

LL −2289.63 −1084.49 −1552.74 −1993.37 −835.36 −742.49 −1092.35BIC 4633.60 2223.31 3159.82 4041.08 1725.05 1539.32 2239.04

Jarque Bera 4102.8546∗∗∗ 141.7671∗∗∗ 287.3423∗∗∗ 23.7308∗∗∗ 569.9882∗∗∗ 1308.3212∗∗∗ 19.8160∗∗∗

Ljung Box (25) 48.0789∗∗∗ 52.0254∗∗∗ 46.0625∗∗∗ 12.0729 27.0364 23.9060 17.7446ARCH (25) 51.4380∗∗∗ 26.1708 34.5267∗ 18.9056 48.9345∗∗∗ 33.5821 21.4819

Table A.11: Estimation results from FIAPARCH model with n = 891 observations (Aug 1, 2014-Dec 31,2017). Statistically significant parameters are indicated with asterisk *, **, *** for 10%, 5%, and 1%level of significance.

CRIX Gold Silver WTI S&P 500 MSCI World MSCI EM50

θ0 0.3638∗∗∗ −0.0103 0.0048 −0.1275∗∗∗ 0.0523∗∗∗ 0.0297∗ 0.0196θ1 −0.0268∗∗∗ −0.0065 −0.0312 −0.1087∗∗∗ −0.0996∗∗∗ 0.0662∗∗ 0.1996∗∗∗

ω 0.3807∗∗ 0.1658 1.0860 0.4569∗ 0.2518∗∗∗ 0.1464∗∗∗ 0.2278∗∗∗

α 0.0193 0.2635 0.4377 0.0778 0.0011∗∗∗ 0.0383 0.0000d 0.7712∗∗∗ 0.2651 0.1245 0.8380∗∗ 0.9776∗∗∗ 0.5146∗∗∗ 0.3114∗∗∗

β 0.7042∗∗∗ 0.5286 0.5278 0.9154∗∗∗ 0.8296∗∗∗ 0.4738∗∗ 0.2914∗∗∗

γ −0.2504∗ −0.1291 −0.2770 0.9125∗∗∗ 0.9701∗∗∗ 0.8704∗∗∗ 0.5702∗∗∗

δ 0.2467∗∗ 2.0700∗ 3.0000 0.3692 0.8543∗∗∗ 1.1052∗∗∗ 1.2969∗∗∗

ν 2.3793∗∗∗ 4.3169∗∗∗ 2.9682∗∗∗ 14.6145 5.1972∗∗∗ 6.1522∗∗∗ 16.4508∗

LL −2282.66 −1086.49 −1552.76 −1997.29 −826.14 −740.58 −1092.82BIC 4626.45 2234.10 3166.65 4055.71 1713.41 1542.29 2246.77

Jarque Bera 2807.9919∗∗∗ 135.6287∗∗∗ 364.5237∗∗∗ 44.1969∗∗∗ 532.0118∗∗∗ 2397.7887∗∗∗ 15.3015∗∗∗

Ljung Box (25) 42.4010∗∗ 50.3412∗∗∗ 43.1643∗∗ 11.9732 25.6216 21.7947 17.0946ARCH (25) 62.3583∗∗∗ 19.9310 29.2771 22.3528 13.4596 9.2813 27.0943

08/2014 01/2015 01/2016 01/2017 01/2018

Time

-0.1

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

Weig

hts

Minimum Variance Composition

Gold weights

CRIX weights

Figure A.8: Time-varying weights of minimum-variance portfolios for Gold–S&P 500 and CRIX–S&P 500based on BEKK correlations.

26

Appendix B. Additional Figures

To be made available online.

07/2011 01/2012 01/2013 01/2014 01/2015 01/2016 01/2017 01/2018

Time

-0.4

-0.3

-0.2

-0.1

0

0.1

0.2

0.3

0.4

0.5

0.6

BE

KK

Cor

rela

tion ;

Gold - MSCI World

BEKK ;SG-smoothed ;

Figure B.9: Dynamic correlations of Gold and MSCI World returns obtained with the BEKK-GARCHbetween July 2, 2011 and December 31, 2017, n = 1 695. Unfiltered correlations are plotted in gray,Savitzky-Golay-smoothing is plotted in red. Times of market distress is highlighted in blue.

07/2011 01/2012 01/2013 01/2014 01/2015 01/2016 01/2017 01/2018

Time

-0.4

-0.3

-0.2

-0.1

0

0.1

0.2

0.3

0.4

0.5

0.6

BE

KK

Cor

rela

tion ;

Bitcoin - MSCI World

BEKK ;SG-smoothed ;

Figure B.10: Dynamic correlations of Bitcoin and MSCI World returns obtained with the BEKK-GARCHbetween July 2, 2011 and December 31, 2017, n = 1 695. Unfiltered correlations are plotted in gray,Savitzky-Golay-smoothing is plotted in red. Times of market distress is highlighted in blue.

27

07/2011 01/2012 01/2013 01/2014 01/2015 01/2016 01/2017 01/2018

Time

-0.4

-0.3

-0.2

-0.1

0

0.1

0.2

0.3

0.4

0.5

0.6

BE

KK

Cor

rela

tion ;

Gold - MSCI World and Bitcoin - MSCI World Correlation

Gold - MSCI WolrdBTC - MSCI World

Figure B.11: Smoothed correlations of Bitcoin and Gold returns with MSCI World returns obtained withthe BEKK-GARCH between July 2, 2011 and December 31, 2017, n = 1 695. Times of market distressis highlighted in blue.

Figure B.12: Smoothed correlations of Bitcoin and Gold returns with MSCI EM50 returns obtained withthe BEKK-GARCH between July 2, 2011 and December 31, 2017, n = 1 695. Times of market distressis highlighted in blue.

28

07/2011 01/2012 01/2013 01/2014 01/2015 01/2016 01/2017 01/2018

Time

-0.4

-0.3

-0.2

-0.1

0

0.1

0.2

0.3

0.4

0.5

0.6

BE

KK

Cor

rela

tion ;

Gold - WTI and Bitcoin - WTI Correlation

Gold - WTIBTC - WTI

Figure B.13: Smoothed correlations of Bitcoin and Gold returns with WTI returns obtained with theBEKK-GARCH between July 2, 2011 and December 31, 2017, n = 1 695. Times of market distress ishighlighted in blue.

29